Washington used a 2018 export statute to disable a US-made model for every customer on Earth — and nobody knows where the limit is.
This was the week regulators stopped writing AI rules and started enforcing one nobody saw coming: on June 12 the Commerce Department invoked export-control law to force Anthropic to disable Fable 5 and Mythos 5 worldwide, the first time that authority has reached a model's access rather than a chip. Allies pushed back, 76 cybersecurity veterans called the move dangerous, and a bipartisan group in the House demanded the legal basis in writing. Meanwhile the state-level chatbot wave kept signing into law and the EU AI Act's full-application date crept closer.
Watch & Listen First
The two cleanest playable explainers of the week, both from TechCrunch's Equity desk:
- The US banned Anthropic's Fable 5 release, but the numbers don't seem to care (Jun 19) — the hosts walk the export directive, the jailbreak claim, and whether this is security or politics.
- Is the US government's Anthropic ban accidentally helping the brand? (Jun 19) — a short watch on the strange market reaction to a national-security shutdown.
Key Takeaways
- Export-control law now reaches model access, not just hardware. If your model can be deemed a "military-intelligence" risk, Commerce can order it offline for everyone — including your own foreign-national staff — on a private letter, no rulemaking.
- The legal basis is contested in real time. Lawyers across Lawfare, CSIS and Tech Policy Press agree the ECRA/EAR theory is plausible but unsettled: serving a model from your own servers may not be an "export" at all.
- Allies are treating US AI as a supply risk. The EU Commission, Macron, and Canada's Carney all signaled diversification this week — sovereign-AI demand is now a compliance input, not a slogan.
- State chatbot law is hardening fast. Vermont, Arizona, Rhode Island and others advanced therapy-bot and minor-safety bills in the last seven days; if you ship a wellness or companion product, your obligations now differ by ZIP code.
- Two hard dates are coming: the EU AI Act's full application on Aug 2, 2026, and Colorado's replacement ADMT law on Jan 1, 2027.
The Big Story
The US government orders Anthropic to disable Fable 5 and Mythos 5 for every customer worldwide · Anthropic · Jun 12
-> The directive — which Anthropic attributes to US national-security authorities, and which CSIS traces to the Commerce Department's Bureau of Industry and Security — ordered the company to suspend access for any foreign national anywhere — so the company disabled both models entirely, because it cannot screen users by nationality at the API. The cited hook is a narrow, disputed "jailbreak"; Anthropic says the behavior is widely available from other deployed models and warns the government's standard would "essentially halt all new model deployments for all frontier model providers." For builders, the precedent is the product: a model you license can be switched off by letter, with criminal and civil penalties attached, and the controlling legal theory (ECRA + the EAR's military-intelligence end-use rule) has never been tested on remote model access.
Also This Week
76 cybersecurity veterans sign open letter calling the ban "dangerous" · TechCrunch · Jun 15
-> Alex Stamos, Katie Moussouris, Rachel Tobac and 73 others argue at freefable.org that pulling the strongest models from defenders — while adversaries advance — is the opposite of a national-security win, putting the security community on record against the directive.
EU Commission says the restrictions "should not be discriminatory," opens consequences review · Euronews · Jun 14
-> Brussels is formally assessing what the cutoff means for European users and signaling its own AI and cybersecurity law as a fallback — the first sign a major market may treat unilateral US export action as grounds for regulatory divergence.
South Korea's science ministry signs an AI safety and cybersecurity MOU with Anthropic · Anthropic · Jun 17
-> Even as the export ban cut Korean access to the top models, Seoul's MSIT and the Korea AI Safety Institute agreed to joint Korean-language safety evaluations and cyber-threat intelligence sharing — government-to-government AI governance proceeding around the directive, not because of it.
The ban is straining US alliances, with Canada and France signaling diversification · Al Jazeera · Jun 19
-> Macron called the reaction "strictly nationalist" and Canada's Carney warned that accepting single-vendor, single-country dependence would itself be the mistake — for multinationals, model sourcing is now a geopolitical-risk line item.
State chatbot and therapy-AI laws keep landing · Transparency Coalition · Jun 18
-> Vermont's H.816 therapy-chatbot ban (signed Jun 17) and H.211 data-broker law (Jun 16), Arizona's HB 2311 minor-safety bill, and Rhode Island's H.7349 advanced in the last week — anyone deploying a companion, wellness, or minor-facing assistant now faces a thickening, state-by-state patchwork.
Deadlines & Compliance
- Aug 2, 2026: The EU AI Act reaches full applicability (European Commission) — the headline milestone when the Act broadly applies to providers placing AI on the EU market; its general-purpose-AI rules already applied from Aug 2, 2025.
- Jan 1, 2027: Colorado's repealed-and-replaced AI law (SB 26-189) takes effect, swapping the original algorithmic-discrimination duty for a narrower automated-decision-making disclosure-and-rights regime — re-scope any Colorado compliance program built for the old SB 24-205.
- Now: If you serve frontier-model access to foreign nationals, document your access-control posture; the Anthropic directive shows BIS can act on a private letter with immediate effect.
Worth Reading
- A Kill Switch for Frontier AI (Lawfare, Jun 15) — the sharpest legal read on why this is "the first time that export controls have been enforced to control access to an AI model," and why that may need Congress.
- Did the US Government Just Set an AI Export Precedent by Blocking Mythos? (Tech Policy Press, Jun 15) — frames the choice between an incremental-risk and a capability-based control theory, and warns the precedent is crystallizing through ad hoc enforcement rather than rulemaking.
- The Department of Commerce Restricted Access to Anthropic's Latest Models. What Comes Next? (CSIS, Jun 16) — maps the three realistic exits (negotiated retraction, litigation, or onerous identity verification) and the risk of pushing customers to Chinese or open-weight alternatives.
The rulebook didn't change this week — the enforcer found a new switch, and every frontier deployer is now downstream of a letter.